fix: replace words wrapped in leading punctuation

[NishchayMahor]

What

tokenize.utils.replace_words handles a word's trailing punctuation but not its leading punctuation, so a word wrapped in quotes/parens/brackets or with a leading dash is never replaced — while the same word with trailing punctuation is:

replace_words(text="world!",   replacements={"world": "universe"})  # -> "universe!"   ✅
replace_words(text="(world)",  replacements={"world": "universe"})  # -> "(world)"     ❌ should be "(universe)"
replace_words(text='"world"',  replacements={"world": "universe"})  # -> '"world"'     ❌
replace_words(text="-world",   replacements={"world": "universe"})  # -> "-world"      ❌

split_words(ignore_punctuation=False) returns tokens with their surrounding punctuation ("(world)", '"world"', "-world"), but _process_words computes no_punctuation = word.rstrip(PUNCTUATIONS) — trailing only — so the dict lookup misses whenever there's leading punctuation. Asymmetric and surprising for TTS pronunciation substitution.

Fix

In _process_words (shared by both the sync and async paths), strip punctuation on both sides for the lookup and preserve whatever punctuation surrounded the word:

core = word.strip(punctuations)
leading_off  = len(word) - len(word.lstrip(punctuations))
trailing_off = len(word) - len(word.rstrip(punctuations))
...
text = text[: start + offset + leading_off] + replacement + text[end + offset - trailing_off :]
offset += len(replacement) - len(core)

This reduces exactly to the previous logic when leading_off == 0, so existing behavior (trailing punctuation, the A.B.C→A.B.C.D internal-dot case) is unchanged.

Testing

Added test_replace_words_leading_punctuation to tests/test_tokenizer.py covering (world), "world", -world, hello (world)., trailing-only, and a no-false-positive case (worldly). The existing test_replace_words still passes (verified). Pure string functions — offline.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.